Feature Selection from Normal Operation Data in Nuclear Power Plants Using ShapeBased Clustering

The application of artificial intelligence in nuclear power plants is often hindered by the high dimensionality of sensor data and the scarcity of labeled records for abnormal events. While most existing diagnostic models rely on supervised learning with simulated fault data, this study proposes an unsupervised feature selection framework that leverages the abundant data available from normal operations. We hypothesize that the tight physical coupling within nuclear power plant systems causes redundant variables to exhibit synchronized time-series shapes, even during normal load-following transients. To identify these redundancies, we introduce a shape-based clustering approach using dynamic time warping, which robustly groups variables with similar dynamic behaviors regardless of temporal delays. This method is evaluated against traditional trend-based and agglomerative clustering techniques using data from a large-scale generic pressurized water reactor simulator. From an initial set of thousands of variables, the proposed method successfully identified a minimal subset of representative variables, achieving an extensive dimensionality reduction. The validity of the selected features was verified by training a simplified multi-layer perceptron to diagnose several distinct abnormal scenarios. The model trained on the reduced feature set demonstrated diagnostic accuracy comparable to a reference model using all variables, confirming that the selected representatives preserve the essential dynamic characteristics of the plant. This study demonstrates that unsupervised analysis of normal operation data can effectively filter redundant information, enabling the development of lightweight, interpretable, and computationally efficient diagnostic systems suitable for real-time monitoring.

Keywords: Normal operation, clustering, feature selection, dynamic time warping.